ComfyUI-Zegami

ComfyUI-Zegami

Push every ComfyUI generation — image or video — into a Zegami collection for visual triage, comparison, and curation.

Generating thousands of variations and triaging them by scrolling a folder? Drop the Zegami Batch Export node onto any workflow and every output lands in a Zegami collection — a filterable visual grid with similarity search and UMAP clustering. Surface the best 50 of 5,000, audit a LoRA training set for duplicates, compare samplers/CFG/seeds across a grid, or publish a shareable gallery as the client deliverable.

Install

ComfyUI Manager (recommended): search for "Zegami" and install.

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/zegami/comfyui-zegami
pip install -r comfyui-zegami/requirements.txt   # just `requests`

Video export needs ffmpeg on your PATH (a system binary, not a pip package).

Authenticate

Create a collection-scoped API key in Zegami (Collection → Settings → API access) — it can only write to that one collection, so it's safe to keep in a shared workflow file. Provide it in priority order:

1. ZEGAMI_API_KEY environment variable (recommended)
2. ~/.zegami/config.json: { "api_key": "zeg_…", "endpoint": "https://…" }
3. the node's api_key_override input (last resort)

The key is never read from or written into a workflow JSON — published workflows stay credential-free.

The node

| Input | Notes | |---|---| | images | Standard ComfyUI IMAGE batch — the primary path: each item in the batch is a separate still. | | video | Frames tensor [frames,h,w,c] (Wan / Hunyuan / LTX / Mochi) → one MP4, or a VHS_VIDEO passthrough. | | collection_id | Target Zegami collection. | | tags, notes | Free-form, attached to every output. | | fps | Frame rate for video encoding (default 16). | | api_key_override, endpoint_override | Fallbacks; prefer env / config. | | enabled | Toggle upload off during iteration without removing the node. |

Outputs: images (pass-through, so it sits inline before a Save node) and upload_status (JSON for chaining a notification node).

The node captures the full prompt graph + workflow JSON + execution metadata as one opaque blob. In Zegami, the Calculated Columns → From JSON path tool (with one-click smart defaults for KSampler / LoraLoader / CheckpointLoader / CLIPTextEncode) turns prompt.6.inputs.seed into a Seed column, etc. — no need for the node to keep pace with every node type.

How it works

Per batch the node: encodes media locally first (the fail-soft anchor) → builds a zip + a metadata.csv (with a _comfy_json column) → stages + enqueues via Zegami's ingest contract on a background queue. A Zegami failure never breaks your generation — on a permanent failure it writes a .zegami-pending sidecar next to the media.

| Scenario | Behaviour | |---|---| | No API key found | Save locally, return error status. Generation completes. | | Network failure | Retry with backoff; then a .zegami-pending sidecar. | | Invalid collection / 4xx | Clear error in upload_status. Generation completes. | | Video encode fails | Upload the first frame as a still instead. | | Partial metadata | Upload whatever was captured — never fail on metadata. |

Status & roadmap

• Images: fully supported end-to-end today.
• Video: grid thumbnails + badges work via the poster + media_kind/duration columns. Full in-app video playback (inspector + hover preview) needs the original MP4 served from the collection's raw_assets/, which the current zip-ingest path doesn't populate — tracked as a follow-up (an AARO-style video ingest or a dedicated raw-asset upload step).
• v1.1: .zegami-pending retry CLI; in-node progress.
• v1.2: a ZegamiCollectionInput node to pull triaged images back into a workflow.
• v1.3: deeper smart-flattening learned from real workflows.

MIT licensed. Built by Zegami.